# Can the Arc A580 run UserLM-2.0?

> Yes. The Arc A580 runs UserLM-2.0 fully in VRAM at Q5_K_M (about 7.1 GB of 8.0 GB), at roughly 49 tokens/s.

Arc A580: 8 GB (8 GB usable), 512 GB/s. Estimates at an 8K context.

## Every quantisation

| Quant | Quality | ≈ GB | Fit | Est. tok/s |
|---|---|---|---|---|
| Q2_K | Heavy loss | 4.08 | fits | 107 |
| Q3_K_M | Noticeable loss | 5.34 | fits | 72 |
| Q4_K_M | Small loss | 6.31 | fits | 58 |
| Q5_K_M (pick) | Near-lossless | 7.11 | fits | 49 |
| Q6_K | Near-lossless | 7.95 | tight | 43 |
| Q8_0 | Lossless | 9.85 | offload | 14 |
| F16 | Lossless | 17.21 | offload | 3.5 |

## Run it

OpenAI-compatible server with a web UI (llama.cpp):

```bash
llama-server -hf liodon-ai/UserLM-2.0-imatrix-GGUF:Q5_K_M
```

Docker with Vulkan (llama.cpp):

```bash
docker run -d --name llama --device /dev/dri -p 8080:8080 \
  -v ~/.cache/llama.cpp:/root/.cache/llama.cpp \
  ghcr.io/ggml-org/llama.cpp:server-vulkan \
  -hf liodon-ai/UserLM-2.0-imatrix-GGUF:Q5_K_M --host 0.0.0.0 --port 8080 -ngl 99
```

Chat with the pipeline API (transformers):

```python
from transformers import pipeline

pipe = pipeline("text-generation", model="microsoft/UserLM-2.0", dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Explain self-hosting in one paragraph."}]
out = pipe(messages, max_new_tokens=512)
print(out[0]["generated_text"][-1]["content"])
```


## Chat models that fit comfortably

- [meta-llama/Llama-3.1-8B-Instruct](https://selfhost.directory/models/meta-llama/llama-3.1-8b-instruct/on/arc-a580): Q5_K_M ≈7.11 GB, est. 49 tok/s
- [Qwen/Qwen3-8B](https://selfhost.directory/models/qwen/qwen3-8b/on/arc-a580): Q5_K_M ≈7.35 GB, est. 48 tok/s
- [meta-llama/Meta-Llama-3-8B-Instruct](https://selfhost.directory/models/meta-llama/meta-llama-3-8b-instruct/on/arc-a580): Q5_K_M ≈7.11 GB, est. 49 tok/s
- [Qwen/Qwen2.5-7B-Instruct](https://selfhost.directory/models/qwen/qwen2.5-7b-instruct/on/arc-a580): Q6_K ≈7.05 GB, est. 45 tok/s
- [xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic](https://selfhost.directory/models/xtools-at/deepseek-v4-pro-qwen3.5-9b-uncensored-heretic/on/arc-a580): Q4_K_M ≈7.15 GB, est. 49 tok/s
- [mistralai/Mistral-7B-Instruct-v0.2](https://selfhost.directory/models/mistralai/mistral-7b-instruct-v0.2/on/arc-a580): Q6_K ≈7.33 GB, est. 47 tok/s

## Other GPUs that run it

- [GeForce RTX 4060](https://selfhost.directory/models/microsoft/userlm-2.0/on/rtx-4060): Q5_K_M, est. 26 tok/s
- [Apple M4 (16 GB)](https://selfhost.directory/models/microsoft/userlm-2.0/on/apple-m4-16gb): Q8_0, est. 7.7 tok/s
- [Radeon RX 6700 XT](https://selfhost.directory/models/microsoft/userlm-2.0/on/rx-6700-xt): Q8_0, est. 25 tok/s
- [GeForce RTX 3060 12GB](https://selfhost.directory/models/microsoft/userlm-2.0/on/rtx-3060): Q8_0, est. 23 tok/s
- [Arc B580](https://selfhost.directory/models/microsoft/userlm-2.0/on/arc-b580): Q8_0, est. 29 tok/s
- [Tesla T4](https://selfhost.directory/models/microsoft/userlm-2.0/on/tesla-t4): Q8_0, est. 21 tok/s
- [Arc A770 16GB](https://selfhost.directory/models/microsoft/userlm-2.0/on/arc-a770): Q8_0, est. 36 tok/s
- [Radeon RX 7800 XT](https://selfhost.directory/models/microsoft/userlm-2.0/on/rx-7800-xt): Q8_0, est. 40 tok/s

---
Model: https://selfhost.directory/models/microsoft/userlm-2.0. GPU: https://selfhost.directory/models/gpus/arc-a580. Source: https://selfhost.directory/models/microsoft/userlm-2.0/on/arc-a580